YOLOV5 network architecture

Overall architecture

• Backbone(Main):Focus, BottleneckCSP, CSP
• Head : PANet + Dectect (Yolov3/Yolov4 Head)

Code part

# parameters
nc: 80  # NUMBER OF CLASSES Category Number
depth_multiple: 0.33  # Model Depth Multiple Control Model Depth (BottleneckCSP)
width_multiple: 0.50  # LAYER CHANNEL MULTIPLE controls the number of conv channels (number of convolutionary nuclear)
# depth_multiple The layer scaling factor of the BottleneckCSP module, multiplying all the Bottleneck's Bottleneck's Bottleneck to get the final number.
#width_multiple Indicates the scaling factor of the convolution channel, which is the setting of the Backbone and Head section inside the Conv channel, all multiplied by this coefficient.
# The model design of different complexity can be achieved by these two parameters.

# anchors
anchors:   #Prick box or anchor frame
  - [10,13, 16,30, 33,23]  # P3 / 8 Represents Stride = 8 is also 8 times the size of the size after the sampling scale is 10 lengths 13
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

# YOLOv5 backbone
backbone:
  # [from, number, module, args]
  # From column parameters: Current module input from which layer output; -1 represents the input obtained from the previous layer
  # Number column parameters: This module is repeated; 1 means only one, 3 indicates three identical modules
  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 1-p2 / 4; 128 Indicates 128 volume nucleus, 3 represents 3 x 3 of the convolutionary core, 2 indicates that the step length is 2.
   [-1, 3, BottleneckCSP, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
   [-1, 9, BottleneckCSP, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
   [-1, 9, BottleneckCSP, [512]],
   [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 3, BottleneckCSP, [1024, False]],  # 9
  ]

# YOLOv5 head
# The author does not divide the NECK module, so the head section contains the PaNet + Detect section.
head:
  [[-1, 1, Conv, [512, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 3, BottleneckCSP, [512, False]],  # 13

   [-1, 1, Conv, [256, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 3, BottleneckCSP, [256, False]],  # 17 (P3/8-small)

   [-1, 1, Conv, [256, 3, 2]],
   [[-1, 14], 1, Concat, [1]],  # cat head P4
   [-1, 3, BottleneckCSP, [512, False]],  # 20 (P4/16-medium)

   [-1, 1, Conv, [512, 3, 2]],
   [[-1, 10], 1, Concat, [1]],  # cat head P5
   [-1, 3, BottleneckCSP, [1024, False]],  # 23 (P5/32-large)

   [[17, 20, 23], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]

Number of x ___multiple is the final number

The network is constantly deepening (how much is used), is constantlyIncrease network feature extractionwithFeature fusionAbility.

How much is the number of convolution, and it also affects the third dimension of the feature map of the convolution, iethickness, Big white here is the networkwidth。

FOCUS

It is equivalent to twice the sample, the quantity is four times,Can reduce floating point calculations but will not increase speed, do not affect MAP

[b,c,h,w]------->[b,c*4,h//2,w//2]

Where RES Block is the residual module

Pool

You can increase your field and extract the most important characteristics but will not reduce the amount of operation.

After the filling of the above poolization operation, the feature is not reduced.

PANet

That is, when the picture is only sampled, the original picture is not good and the picture after the sample is characterized, in order to solve this problem.Another line green lineAfter the upper sampling, the feature map can be fused on the same scale, and finallySplicing

So get information is more rich